AlphaFold

Every two years since 1994, structural biologists have run a competition called CASP. Organisers take proteins whose shapes have been determined experimentally but not yet published, release only the sequences, and let teams predict. The real answers are held back, so nobody can tune to them.

It is the ImageNet arrangement, in biology. A shared problem, hidden answers, everyone measured the same way.

In 2020, at CASP14, a DeepMind system called AlphaFold2 produced high-accuracy structures for 87 of 92 assessed . For 58, its best prediction reached a score the assessors described as comparable with experimental accuracy.

The assessors also combined performance across targets into one ranking. AlphaFold2 scored 244.0; the next group scored 90.8. Higher was better. This was a large measured gap under the same hidden test.

AlphaFold depended on experimental structures deposited in the Protein Data Bank, along with large databases of protein sequences. Its training data therefore carried decades of laboratory work.

It was not reinforcement learning in the AlphaGo sense. The same organisation used a different learning arrangement because this problem came with experimentally determined structures that could act as targets.

In 2024, Demis Hassabis and John Jumper shared half the Nobel Prize in Chemistry for protein structure prediction. The other half went to David Baker for computational protein design.